ON BELLMAN SYSTEMS WITHOUT ZERO ORDER TERM IN THE

advertisement
126 (2001)
MATHEMATICA BOHEMICA
No. 2, 275–280
ON BELLMAN SYSTEMS WITHOUT ZERO ORDER TERM IN THE
CONTEXT OF RISK SENSITIVE DIFFERENTIAL GAMES
A. Bensoussan, Paris, J. Frehse, Bonn
Dedicated to Prof. J. Nečas on the occasion of his 70th birthday
Abstract. Bellman systems corresponding to stochastic differential games arising from a
cost functional which models risk aspects are considered. Here it leads to diagonal elliptic
systems without zero order term so that no simple L∞ -estimate is available.
Keywords: diagonal elliptic systems, quadratic growth, stochastic differential games,
Bellman equation, risk sensitive control
MSC 2000 : 91A15, 93C20, 93E20, 91A80
1. Introduction
Bellman systems of the type, say
(1.1)
1
− ∆uν + αuν = Hν (x, Du), ν = 1, . . . , N
2
uν |∂O = 0
with certain structure conditions on H and quadratic cost functionals have been
studied in order to solve stochastic games.
For instance, in [1] the authors solved the differential game
(1.2)
dy = g(y(t)) +
N
vµ (t) dt + dw(t),
µ=1
y(0) = x, x ∈
n
,
where v1 (·), . . . , vN (·) are controls at the disposal of N decision makers. In (1.2),
w(t) is a Wiener process in n , and yx,v = y(·) is the solution of an Ito stochastic
275
differential equation. Let O be an open smooth bounded domain of
n
, and let
/ O}
τ = inf{t ; yx,v (t) ∈
(1.3)
be the first exit time of the process yx,v (t) outside O. Using the notation
v(t) = (vν (t), v ν (t))
(1.4)
where v ν represents the vector of all components which are different from vν and
(1.5)
v ν (t) =
vµ (t), ν, µ = 1, . . . , N,
µ=ν
we consider the cost function of the player ν, given by
(1.6)
Jν (x, v(·)) = Jν (x, vν (·), v ν (·))
τ
=
e−αt fν (yx,v (t)) + 12 |vν (t)|2 + θvν (t) · v ν (t) dt.
0
A Nash point of the functionals Jν (x, v(·)) is a control v̂(·) such that
(1.7)
Jν (x, v̂ν (·), v̂ ν (·)) Jν (x, vν (·), v̂ ν (·)), ν = 1, . . . , N
for any admissible control v(·) = (v1 (·), . . . , vN (·)). Defining a function
(1.8)
Lν (v, p) =
1
|vν |2 + θvν · v ν + pν ·
vµ
2
µ
where p = (p1 , . . . , pN ) ∈ nN , v = (v1 , . . . , vN ) ∈ nN and considering a Nash
point v̂1 (p), . . . , v̂N (p) of the functions (1.8) (the definition is similar to (1.7), but it
is pointwise in x), then setting
Lν (p) = Lν (v̂(p), p)
(1.9)
it is proved that the functions
(1.10)
uν (x) = Jν (x, v̂(·))
are solutions of the system of partial differential equations
(1.11)
276
1
− ∆uν − g(·) · Duν + αuν = fν + Lν (Du)
2
which is of the form (1.1) with
(1.12)
Hν (x, p) = Lν (p) + fν (x) + g(x) · pν .
Note that the discount factor e−αt gives the 0-order term αuν in the Bellman system
(1.11).
This term helps very much in obtaining L∞ -estimates, via Maximum Principle
type of argument.
In recent years, there has been a rising interest in taking into consideration risk
aspects in the cost functions. One convenient way of modelling risk is to consider
the cost functions (instead of (1.6))
(1.13)
Jνδ (x, v(·)) = Jνδ (x, vν (·), v ν (·))
τ 1
1
= log exp δ
fν (yx,v (t)) + |vν (t)|2 + θvν (t) · v ν (t) dt
δ
2
0
where δ is called the risk factor (δ > 0 represents an aversion to risk, δ < 0 represents
an attraction to risk).
τ
Note that in the integral , there is no discount factor any more.
0
The reason for omitting the discount factor is that Nash points of functionals of
the type (1.13) are amenable to systems of partial differential equations similar to
(1.11). Introducing the discount factor leads unfortunately to parabolic systems and
not to elliptic ones.
If v̂(·) is a Nash point for (1.13), then
uν (x) = Jν (x, v̂(·))
is a solution of the system
(1.14)
1
δ
− ∆uν − g(x) · Duν = |Duν |2 + fν (x) + Lν (Du)
2
2
and thus we are led to systems of the type
(1.15)
1
− ∆uν = Hν (x, Du)
2
uν |∂O = 0.
One of the main difficulties is to recover L∞ -estimates. In this note we present some
cases where the L∞ -estimate is available. In particular, we show that the drift g(x)
can have an influential role in obtaining these estimates.
277
2. Statement of problem and results
2.1. Assumptions and model. We consider here the system
1
− ∆uν − g(x) · Duν = Hν (x, Du)
2
uν |∂O = 0,
(2.1)
where Hν (x, p) are Carathéodory functions, g ∈ W 1,∞ (O) with the following assumptions
(2.2)
Hν (x, p) −λ
∀x, p,
ν
Hν (x, p) λν + λ0ν |pν |2 .
(2.3)
If Γ is an N × N -matrix and if we set
HνΓ (x, p) = (ΓH)ν (x, Γ−1 p)
(2.4)
where H(x, p) represents the vector (H1 (x, p), . . . , HN (x, p)) then we assume that
(2.5)
there exists a matrix Γ such that HνΓ (x, p) = Q(x, p) · pν + Hν0 (x, p)
with
|Q(x, p)| k + K|p|,
|Hν0 (x, p)| kν + Kν
|pµ |2 .
(2.6)
(2.7)
µν
The assumptions (2.5), (2.6), (2.7) represent the special structure assumption (note
that this special structure may not be available on the original Hν but only after a
linear manipulation represented by the matrix Γ).
An additional smallness condition on the product λν λ0ν is assumed, namely
4λν λ0ν < k0 + inf div g
(2.8)
where k0 is the constant arising in the Poincaré inequality
(2.9)
k0 ϕ2 dx |Dϕ|2 dx
∀ϕ ∈ H01 (O).
O
O
2.2. Statement of the results.
Theorem 2.1. Assuming O to be smooth bounded and Hν (x, p), Carathéodory
functions satisfying (2.2), (2.3), (2.5), (2.6), (2.7) as well as (2.8) there exists a
solution u of (2.1) such that u ∈ (W 2,s (O))N , for every 2 s < ∞.
278
3. Proof of the L∞ -estimate
We will not give the complete proof of Theorem 2.1, nevertheless, we will show
some details how the L∞ -estimates are obtained.
Write
ũ =
uν ,
then adding up the equations (2.1) we have
1
− ∆ũ − gDũ −λ.
2
(3.1)
For any point ξ of O consider the Green function
1
− ∆Gξ + div(gGξ ) = δ(x − ξ)
2
Gξ |∂O = 0.
(3.2)
We test (3.1) with ũ− Gξ obtaining, from the definition of the Green function,
1 −
1
− 2
ξ
2
D(ũ ) DG dx + (ũ (ξ)) λ ũ− Gξ dx.
(3.3)
4
2
O
O
Suppose now ξ is a point where ũ− reaches a positive maximum (necessarily in O),
then we get
ũ− ∞ 2λ Gξ dx C
O
∞
so that we have proved the first L -estimate
(3.4)
uν −C.
Next, we introduce the function Eν = exp 2λ0ν uν . We can check from (2.1) and
assumption (2.3) that
1
− ∆Eν − gDEν 2λν λ0ν Eν .
2
(3.5)
Testing (3.5) with (Eν − 1)+ , which vanishes on the boundary, yields
|D(Eν − 1)+ |2 dx +
div g(Eν − 1)+2 dx
O
O
4λν λ0ν
+2
(Eν − 1)
O
dx +
4λν λ0ν
(Eν − 1)+ dx
O
279
and from Poincaré’s inequality we obtain
(k0 + div g)(Eν − 1)+2 dx 4λν λ0ν (Eν − 1)+2 dx + 4λν λ0ν (Eν − 1)+ dx.
O
O
O
Thanks to the smallness condition (2.8), we deduce easily
E 2 dx C.
(3.6)
O
Using this knowledge we are going to check that E is in L∞ , without using anymore
the smallness condition. For that purpose, we test again (3.5) with Eν Gξ , using the
Green function (3.2). We obtain
1 2
0
(3.7)
E (ξ) − 1 2λν λν Eν2 Gξ dx,
2 ν
O
hence, taking ξ as a point of maximum of Eν2 ,
2
0 2
ξ
0
2
(3.8)
Eν ∞ 1 + 4λν λν L
G dx + 4λν λν Eν ∞
O
But from (3.6) one has
Meas {Eν > L} and thus
{Eν >L}
Gξ dx
{Eν >L}
∀L.
C
,
L2
Gξ dx CGξ Lq
1
.
L2q
So by picking L sufficiently large, we can make the coefficient of Eν2 ∞ on the right
hand side of (3.8) as small as we wish, in particular, strictly smaller than 1. So (3.8)
yields an estimate on Eν2 ∞ .
3.1. We see from (2.8) that, if inf div g is large, the limitation on the
product λν λ0ν is not so restrictive. The role of the drift, as a way to soften some
restrictions, has already been investigated by H. Nagaı̈ [2]. Furthermore, if λν 0,
λ0ν may be “large”.
References
[1] A. Bensoussan, J. Frehse: Nonlinear elliptic systems in stochastic game theory. J. Reine
Angew. Math. Mathematik 350 (1984), 23–67.
[2] H. Nagaı̈: Bellman equation of risk sensitive control. SIAM J. Control Optim. 34 (1996),
74–101.
Authors’ addresses: A. Bensoussan, CNES, 2, Place Maurice Quentin, 75039 Paris
CEDEX 01, France; J. Frehse, Institut für Angewandte Mathematik, Universität Bonn,
Beringstr. 6, 53115 Bonn, Germany, e-mail: aaa@iam.uni-bonn.de.
280
Download